Merge branch 'develop' into core_add_fc_op

doc/howto/dev/new_op_cn.md

@@ -169,6 +169,8 @@ class MulKernel : public framework::OpKernel {
 `MulKernel`需要重写`Compute`接口，该接口参数为`const framework::ExecutionContext& context`, `ExecutionContext`相比`InferShapeContext`增加了设备类型，同样可获取到输入输出和属性参数，`Compute`函数里写具体实现时。
    
 注意，不同设备(CPU、GPU)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。`MulOp`的CPU、GPU实现共享同一个`Kernel`，`OpKernel`不共享的例子可以参考[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 
+
+为了使得`OpKernel`的计算过程书写较为简单，CPU、GPU的代码可以复用，我们通常借助Eigen unsupported Tensor模块来实现。关于在paddle中如何使用Eigen库，请参考对应的使用[文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md)
    
 到此前向Op实现完成，需要在`.cc`文件中注册该op和kernel。反向Op类的定义和Kernel定义与前向Op类似，这里不再重复。但注意，反向Op没有`ProtoMaker`。
    
@@ -188,9 +190,12 @@ REGISTER_OP_CPU_KERNEL(mul_grad,
   - `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op。
   - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulKernel`类。
 
-在 `.cu`文件中注册GPU Kernel。
+在 `.cu`文件中注册GPU Kernel。请注意，如果GPU Kernel的实现是基于Eigen unsupported模块，那么在 `.cu`的最前面请加上宏定义 `#define EIGEN_USE_GPU`
    
 ```c++
+// if use Eigen unsupported module before include head files
+#define EIGEN_USE_GPU
+
 namespace ops = paddle::operators;
 REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
 REGISTER_OP_GPU_KERNEL(mul_grad,
@@ -286,28 +291,50 @@ class TestMulOp(unittest.TestCase):
 
 反向Op单测继承自`GradientChecker`，而`GradientChecker`集成自`unittest.TestCase`，所以反向单测函数需要`test_`开头。
 
- ```
- class MulGradOpTest(GradientChecker):
-    def test_mul(self):
-        op = create_op("mul")
-        inputs = {
+```
+class TestMulGradOp(GradientChecker):
+    def setUp(self):
+        self.op = create_op("mul")
+        self.inputs = {
             'X': np.random.random((32, 84)).astype("float32"),
             'Y': np.random.random((84, 100)).astype("float32")
         }
-        self.compare_grad(op, inputs)      
+
+    def test_cpu_gpu_compare(self):
+        self.compare_grad(self.op, self.inputs)
+
+    def test_normal(self):
         # mul op will enlarge the relative error
         self.check_grad(
-            op, inputs, set(["X", "Y"]), "Out", max_relative_error=0.5)
- ```
+            self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.5)
+
+    def test_ignore_x(self):
+        self.check_grad(
+            self.op,
+            self.inputs, ["Y"],
+            "Out",
+            max_relative_error=0.5,
+            no_grad_set={"X"})
+
+    def test_ignore_y(self):
+        self.check_grad(
+            self.op,
+            self.inputs, ["X"],
+            "Out",
+            max_relative_error=0.5,
+            no_grad_set={"Y"})
+```
+
+下面解释一些关键的地方:
 
    - 调用`create_op("mul")`创建反向Op对应的前向Op。
-   - 定义输入`inputs`。
    - 调用`compare_grad`函数对比CPU、GPU计算结果。
-   - 调用`check_grad`检查梯度稳定性，这里采用数值法检测梯度正确性。
-      - 第一个参数`op` : 前向op。
-      - 第二个参数`inputs` : 输入词典，词典的Key和`ProtoMaker`定义保持一致。
-      - 第三个参数`set(["X", "Y"])` : 指定对输入变量`X`、`Y`做梯度检测。
+   - `test_normal`中调用`check_grad`检查梯度稳定性，这里采用数值法检测梯度正确性。
+      - 第一个参数`self.op` : 前向Op。
+      - 第二个参数`self.inputs` : 输入词典，词典的Key和`ProtoMaker`定义保持一致。
+      - 第三个参数`["X", "Y"]` : 指定对输入变量`X`、`Y`做梯度检测。
       - 第四个参数`"Out"` : 指定前向网络最终的输出目标变量`Out`
+   - `test_ignore_x`和`test_ignore_y`分支测试只需要计算一个输入梯度的情况。
 
 
 ### 编译和执行 

doc/howto/dev/use_eigen_cn.md

+## 在Paddle中如何使用Eigen
+
+神经网络本质上是一个计算图，计算需要的数据存放在`Tensor`中，而计算过程是由`Operartor`来描述的。在执行时，`Operator`调用对应`OpKernel`中的`Compute`接口，实现对`Tensor`的操作。
+
+
+### Eigen Tensor模块
+
+Eigen Tensor模块对element-wise计算提供了强大的支持，并且书写一份代码，可以同时在CPU、GPU执行。但Eigen Tensor是一个正在开发中的模块，因此可能测试不够完备，文档较少。
+
+关于Eigen Tensor模块的详细介绍请参考[文档1](https://github.com/RLovelett/eigen/blob/master/unsupported/Eigen/CXX11/src/Tensor/README.md) 和[文档2](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md)
+
+
+### paddle::framework::Tensor
+
+Paddle Tensor定义在framework目录下，其主要接口如下：
+
+```cpp
+class Tensor {
+ public:
+  /*! Return a pointer to mutable memory block. */
+  template <typename T>
+  inline T* data();
+  
+  /**
+   * @brief   Return a pointer to mutable memory block.
+   * @note    If not exist, then allocation.
+   */
+  template <typename T>
+  inline T* mutable_data(platform::Place place);
+  
+  /**
+   * @brief     Return a pointer to mutable memory block.
+   *
+   * @param[in] dims    The dimensions of the memory block.
+   * @param[in] place   The place of the memory block.
+   *
+   * @note      If not exist, then allocation.
+   */
+  template <typename T>
+  inline T* mutable_data(DDim dims, platform::Place place);
+  
+  /*! Resize the dimensions of the memory block. */
+  inline Tensor& Resize(const DDim& dims);
+  
+  /*! Return the dimensions of the memory block. */
+  inline const DDim& dims() const;
+
+ private:  
+  /*! holds the memory block if allocated. */
+  std::shared_ptr<Placeholder> holder_;
+  
+  /*! points to dimensions of memory block. */
+  DDim dim_;
+};
+```
+
+`Placeholder`的作用是延迟分配内存，即我们可以先定义一个Tensor，然后使用Resize接口设置Tensor的大小，最后再调用mutable_data接口分配实际的内存。
+
+```cpp
+paddle::framework::Tensor t;
+paddle::platform::CPUPlace place;
+// set size first
+t.Resize({2, 3});
+// allocate memory on CPU later
+t.mutable_data(place);
+```
+
+### paddle::framework::Tensor使用样例
+下面以AddOp为例说明Tensor的使用过程：
+
+- InferShape
+
+在运行神经网络计算图时，我们先调用每个`Operator`的`InferShape`接口，根据输入Tensor的大小来设置输出Tensor的大小，`Resize`接口会被调用。
+
+```cpp
+void InferShape(const framework::InferShapeContext &ctx) const override {
+  PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
+                    ctx.Input<Tensor>("Y")->dims(),
+                    "Two input of Add Op's dimension must be same.");
+  ctx.Output<Tensor>("Out")->Resize(ctx.Input<Tensor>("X")->dims());
+}
+```
+
+
+- Run
+
+`Operator`的`Run`接口最终会调用对应`OpKernel`的`Compute`接口，在这时真正的分配内存，`mutable_data`接口会被调用。
+
+```cpp
+void Compute(const framework::ExecutionContext& context) const override {
+  auto* input0 = context.Input<Tensor>("X");
+  auto* input1 = context.Input<Tensor>("Y");
+  auto* output = context.Output<Tensor>("Out");
+
+  output->mutable_data<T>(context.GetPlace());
+
+  auto x = EigenVector<T>::Flatten(*input0);
+  auto y = EigenVector<T>::Flatten(*input1);
+  auto z = EigenVector<T>::Flatten(*output);
+
+  auto place = context.GetEigenDevice<Place>();
+
+  z.device(place) = x + y;
+}
+```
+
+
+### paddle::framework::Tensor到EigenTensor的转换
+
+如上一小节所示，在具体的计算中，我们需要先把输入Tensor和输出Tensor转换为Eigen支持的格式。我们在[eigen.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/eigen.h)中提供了一些全局函数用来实现paddle::framework::Tensor到EigenTensor/EigenMatrix/EigenVector/EigenScalar的转换。
+
+以EigenTensor为例，做一个介绍
+
+```cpp
+Tensor t;
+float* p = t.mutable_data<float>(make_ddim({1, 2, 3}), platform::CPUPlace());
+for (int i = 0; i < 1 * 2 * 3; i++) {
+  p[i] = static_cast<float>(i);
+}
+
+EigenTensor<float, 3>::Type et = EigenTensor<float, 3>::From(t);
+```
+
+From是EigenTensor模板提供的一个接口，可以实现从paddle::framework::Tensor到对EigenTensor的转换。由于Tensor的rank是模板参数，因此在转换时需要显示的指定。
+
+在Eigen中，不同rank的Tensor是不同类型，Vector是rank为1的Tensor。需要额外注意的是，EigenVector<T>::From方法是把paddle中的一维Tensor转为Eigen的一维Tensor，在这里用EigenVector来表示；而EigenVector<T>::Flatten方法是把paddle中的一个Tensor进行reshape操作，压扁成为Eigen的一维Tensor，类型仍然为EigenVector。
+
+更多的转换方法请参考eigen_test.cc中的[单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/eigen_test.cc)。
+
+
+
+### 实现计算
+
+当需要完成计算时，我们需要等式左边的EigenTensor调用device接口。在这里需要注意的是，这里的EigenTensor之间的运算只是改变了原有Tensor中的数据，而不会改变原有Tensor的shape信息。
+
+```cpp
+auto x = EigenVector<T>::Flatten(*input0);
+auto y = EigenVector<T>::Flatten(*input1);
+auto z = EigenVector<T>::Flatten(*output);
+auto place = context.GetEigenDevice<Place>();
+z.device(place) = x + y;
+```
+
+在这段代码中，input0/input1/output可以是任意维度的Tensor。我们调用了EigenVector的Flatten接口，把任意维度的Tensor转为了一维的EigenVector。而在计算结束之后，input0/input1/output的原有shape信息不变。如果想改变原有Tensor的shape信息，可以调用Resize接口进行改变。
+
+由于Eigen Tensor模块的文档较少，我们可以参考TensorFlow的[kernels](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/core/kernels)模块下的相关`OpKernel`的计算代码。

paddle/framework/attribute.cc

@@ -43,6 +43,10 @@ template <>
 AttrType AttrTypeID<std::vector<std::string>>() {
   return STRINGS;
 }
+template <>
+AttrType AttrTypeID<std::vector<std::pair<int, int>>>() {
+  return INT_PAIRS;
+}
 
 Attribute GetAttrValue(const OpDesc::Attr& attr_desc) {
   switch (attr_desc.type()) {
@@ -76,6 +80,14 @@ Attribute GetAttrValue(const OpDesc::Attr& attr_desc) {
       }
       return val;
     }
+    case paddle::framework::AttrType::INT_PAIRS: {
+      std::vector<std::pair<int, int>> val(attr_desc.int_pairs_size());
+      for (int i = 0; i < attr_desc.int_pairs_size(); ++i) {
+        val[i].first = attr_desc.int_pairs(i).first();
+        val[i].second = attr_desc.int_pairs(i).second();
+      }
+      return val;
+    }
   }
   PADDLE_ENFORCE(false, "Unknown OpDesc::AttrDesc::type !");
   return boost::blank();

paddle/framework/attribute.h

@@ -28,7 +28,8 @@ namespace paddle {
 namespace framework {
 
 typedef boost::variant<boost::blank, int, float, std::string, std::vector<int>,
-                       std::vector<float>, std::vector<std::string>>
+                       std::vector<float>, std::vector<std::string>,
+                       std::vector<std::pair<int, int>>>
     Attribute;
 
 typedef std::unordered_map<std::string, Attribute> AttributeMap;

paddle/framework/framework.proto

@@ -22,8 +22,14 @@ enum AttrType {
   INTS = 3;
   FLOATS = 4;
   STRINGS = 5;
+  INT_PAIRS = 6;
 }
 
+message IntPair {
+  required int32 first = 1;
+  required int32 second = 2;
+};
+
 // OpDesc describes an instance of a C++ framework::OperatorBase
 // derived class type.
 message OpDesc {
@@ -37,6 +43,7 @@ message OpDesc {
     repeated int32 ints = 6;
     repeated float floats = 7;
     repeated string strings = 8;
+    repeated IntPair int_pairs = 9;
   };
 
   message Var {

paddle/framework/op_registry_test.cc

@@ -174,36 +174,4 @@ TEST(OpRegistry, CustomChecker) {
   op->Run(scope, dev_ctx);
   int test_attr = op->GetAttr<int>("test_attr");
   ASSERT_EQ(test_attr, 4);
-}
-
+}
\ No newline at end of file
-class TestAttrProtoMaker : public pd::OpProtoAndCheckerMaker {
- public:
-  TestAttrProtoMaker(pd::OpProto* proto, pd::OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddAttr<float>("scale", "scale of test op");
-    AddAttr<float>("scale", "scale of test op");
-  }
-};
-
-TEST(ProtoMaker, DuplicatedAttr) {
-  pd::OpProto op_proto;
-  pd::OpAttrChecker op_checker;
-  auto proto_maker = TestAttrProtoMaker(&op_proto, &op_checker);
-  ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet);
-}
-
-class TestInOutProtoMaker : public pd::OpProtoAndCheckerMaker {
- public:
-  TestInOutProtoMaker(pd::OpProto* proto, pd::OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("input", "input of test op");
-    AddInput("input", "input of test op");
-  }
-};
-
-TEST(ProtoMaker, DuplicatedInOut) {
-  pd::OpProto op_proto;
-  pd::OpAttrChecker op_checker;
-  auto proto_maker = TestInOutProtoMaker(&op_proto, &op_checker);
-  ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet);
-}

paddle/framework/operator_test.cc

@@ -263,4 +263,38 @@ TEST(Operator, Clone) {
   OperatorClone a("ABC", {}, {}, {});
   auto b = a.Clone();
   ASSERT_EQ(a.Type(), b->Type());
+}
+
+class TestAttrProtoMaker : public paddle::framework::OpProtoAndCheckerMaker {
+ public:
+  TestAttrProtoMaker(paddle::framework::OpProto* proto,
+                     paddle::framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddAttr<float>("scale", "scale of test op");
+    AddAttr<float>("scale", "scale of test op");
+  }
+};
+
+TEST(ProtoMaker, DuplicatedAttr) {
+  paddle::framework::OpProto op_proto;
+  paddle::framework::OpAttrChecker op_checker;
+  auto proto_maker = TestAttrProtoMaker(&op_proto, &op_checker);
+  ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet);
+}
+
+class TestInOutProtoMaker : public paddle::framework::OpProtoAndCheckerMaker {
+ public:
+  TestInOutProtoMaker(paddle::framework::OpProto* proto,
+                      paddle::framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("input", "input of test op");
+    AddInput("input", "input of test op");
+  }
+};
+
+TEST(ProtoMaker, DuplicatedInOut) {
+  paddle::framework::OpProto op_proto;
+  paddle::framework::OpAttrChecker op_checker;
+  auto proto_maker = TestInOutProtoMaker(&op_proto, &op_checker);
+  ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet);
 }
\ No newline at end of file

paddle/gserver/layers/Conv3DLayer.cpp

@@ -42,10 +42,10 @@ bool Conv3DLayer::init(const LayerMap &layerMap,
     if (sharedBiases_) {
       CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
       biases_ =
-          std::unique_ptr<Weight>(new Weight(1, numFilters_, biasParameter_));
+          std::unique_ptr<Weight>(new Weight(numFilters_, 1, biasParameter_));
     } else {
       biases_ =
-          std::unique_ptr<Weight>(new Weight(1, getSize(), biasParameter_));
+          std::unique_ptr<Weight>(new Weight(getSize(), 1, biasParameter_));
     }
   }
   return true;
@@ -224,20 +224,31 @@ void Conv3DLayer::bpropData(int i) {
 }
 
 void Conv3DLayer::bpropBiases() {
+  MatrixPtr biases = Matrix::create(biases_->getWGrad()->getData(),
+                                    1,
+                                    biases_->getWGrad()->getElementCnt(),
+                                    false,
+                                    useGpu_);
   MatrixPtr outGradMat = getOutputGrad();
+
   if (this->sharedBiases_) {
-    biases_->getWGrad()->collectSharedBias(*outGradMat, 1.0f);
+    biases->collectSharedBias(*outGradMat, 1.0f);
   } else {
-    biases_->getWGrad()->collectBias(*outGradMat, 1.0f);
+    biases->collectBias(*outGradMat, 1.0f);
   }
 }
 
 void Conv3DLayer::addBias() {
   MatrixPtr outMat = getOutputValue();
+  MatrixPtr bias = Matrix::create(biases_->getW()->getData(),
+                                  1,
+                                  biases_->getW()->getElementCnt(),
+                                  false,
+                                  useGpu_);
   if (this->sharedBiases_) {
-    outMat->addSharedBias(*(biases_->getW()), 1.0f);
+    outMat->addSharedBias(*(bias), 1.0f);
   } else {
-    outMat->addBias(*(biases_->getW()), 1.0f);
+    outMat->addBias(*(bias), 1.0f);
   }
 }
 

paddle/gserver/layers/DeConv3DLayer.cpp

@@ -42,10 +42,10 @@ bool DeConv3DLayer::init(const LayerMap &layerMap,
     if (sharedBiases_) {
       CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
       biases_ =
-          std::unique_ptr<Weight>(new Weight(1, numFilters_, biasParameter_));
+          std::unique_ptr<Weight>(new Weight(numFilters_, 1, biasParameter_));
     } else {
       biases_ =
-          std::unique_ptr<Weight>(new Weight(1, getSize(), biasParameter_));
+          std::unique_ptr<Weight>(new Weight(getSize(), 1, biasParameter_));
     }
   }
   return true;
@@ -191,21 +191,31 @@ void DeConv3DLayer::bpropWeights(int i) {}
 void DeConv3DLayer::bpropData(int i) {}
 
 void DeConv3DLayer::bpropBiases() {
+  MatrixPtr biases = Matrix::create(biases_->getWGrad()->getData(),
+                                    1,
+                                    biases_->getWGrad()->getElementCnt(),
+                                    false,
+                                    useGpu_);
   const MatrixPtr &outGradMat = getOutputGrad();
 
   if (this->sharedBiases_) {
-    biases_->getWGrad()->collectSharedBias(*outGradMat, 1.0f);
+    biases->collectSharedBias(*outGradMat, 1.0f);
   } else {
-    biases_->getWGrad()->collectBias(*outGradMat, 1.0f);
+    biases->collectBias(*outGradMat, 1.0f);
   }
 }
 
 void DeConv3DLayer::addBias() {
   MatrixPtr outMat = getOutputValue();
+  MatrixPtr bias = Matrix::create(biases_->getW()->getData(),
+                                  1,
+                                  biases_->getW()->getElementCnt(),
+                                  false,
+                                  useGpu_);
   if (this->sharedBiases_) {
-    outMat->addSharedBias(*(biases_->getW()), 1.0f);
+    outMat->addSharedBias(*(bias), 1.0f);
   } else {
-    outMat->addBias(*(biases_->getW()), 1.0f);
+    outMat->addBias(*(bias), 1.0f);
   }
 }
 

paddle/operators/fc_op.cc

@@ -31,12 +31,23 @@ class FCOp : public NetOp {
     if (b != framework::kEmptyVarName) {
       AppendOp(framework::OpRegistry::CreateOp(
           "rowwise_add", {{"X", {Output("mul_out")}}, {"b", {Input("b")}}},
-          {{"Out", {Output("mul_out")}}}, {}));
+          {{"Out", {Output("add_out")}}}, {}));
+    } else {
+      AppendOp(framework::OpRegistry::CreateOp(
+          "identity", {{"X", {Output("mul_out")}}},
+          {{"Out", {Output("add_out")}}}, {}));
     }
 
     auto activation = GetAttr<std::string>("activation");
-    AppendOp(framework::OpRegistry::CreateOp(
-        activation, {{"X", {Output("mul_out")}}}, {{"Y", {Output("Y")}}}, {}));
+    if (activation == "identity") {
+      AppendOp(framework::OpRegistry::CreateOp(activation,
+                                               {{"X", {Output("add_out")}}},
+                                               {{"Out", {Output("Out")}}}, {}));
+    } else {
+      AppendOp(framework::OpRegistry::CreateOp(activation,
+                                               {{"X", {Output("add_out")}}},
+                                               {{"Y", {Output("Out")}}}, {}));
+    }
     CompleteAddOp(false);
   }
 };
@@ -49,8 +60,10 @@ class FCOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("W", "The 2D weight matrix of FC operator.");
     AddInput("b", "The 1D bias vector of FC operator");
 
-    AddOutput("Y", "The activated output matrix of FC operator");
-    AddOutput("mul_out", "The non-actived output of FC operator, X * W + b")
+    AddOutput("Out", "The activated output matrix of FC operator");
+    AddOutput("mul_out", "The non-actived output of FC operator, X * W")
+        .AsIntermediate();
+    AddOutput("add_out", "The non-actived output of FC operator, X * W + b")
         .AsIntermediate();
     AddAttr<std::string>("activation", "The activation type of FC operator.")
         .SetDefault("identity")
@@ -65,7 +78,7 @@ learned weights with a matrix multiplication followed by a bias offset
 (optionally).
 
 Equation:
-  Y = Act(sum_n{X_i * W_i} + b)
+  Out = Act(sum_n{X_i * W_i} + b)
 
 where X_i is a 2D matrix of size (M x K), usually M is the minibatch size and
 K is the number of features. W_i is also a 2D matrix of size (K x N),
@@ -78,22 +91,6 @@ Activation type can be set to `identity` (default), `sigmoid` or `softmax`.
   }
 };
 
-class FCGradOp : public NetOp {
- public:
-  FCGradOp(const std::string &type, const framework::VariableNameMap &inputs,
-           const framework::VariableNameMap &outputs,
-           const framework::AttributeMap &attrs)
-      : NetOp(type, inputs, outputs, attrs) {
-    auto y_grad = Input(framework::GradVarName("Y"));
-    auto mul_out_grad = Input(framework::GradVarName("mul_out"));
-    auto x_grad = Output(framework::GradVarName("X"));
-    auto w_grad = Output(framework::GradVarName("W"));
-    auto b_grad = Output(framework::GradVarName("b"));
-
-    CompleteAddOp(false);
-  }
-};
-
 }  // namespace operators
 }  // namespace paddle
 
@@ -104,4 +101,4 @@ USE_OP(sigmoid);
 USE_OP(softmax);
 
 namespace ops = paddle::operators;
-REGISTER_OP(fc, ops::FCOp, ops::FCOpMaker, fc_grad, ops::FCGradOp);
+REGISTER_OP_WITHOUT_GRADIENT(fc, ops::FCOp, ops::FCOpMaker);

paddle/operators/lookup_table_op.h

@@ -30,12 +30,12 @@ class LookupTableKernel : public framework::OpKernel {
     auto ids_t = context.Input<Tensor>("Ids");      // int tensor
     auto output_t = context.Output<Tensor>("Out");  // float tensor
 
-    size_t N = table_t->dims()[0];
-    size_t D = table_t->dims()[1];
+    int N = table_t->dims()[0];
+    int D = table_t->dims()[1];
     auto ids = ids_t->data<int32_t>();
     auto table = table_t->data<T>();
     auto output = output_t->mutable_data<T>(context.GetPlace());
-    for (size_t i = 0; i < product(ids_t->dims()); ++i) {
+    for (ssize_t i = 0; i < product(ids_t->dims()); ++i) {
       PADDLE_ENFORCE_LT(ids[i], N);
       PADDLE_ENFORCE_GE(ids[i], 0);
       memcpy(output + i * D, table + ids[i] * D, D * sizeof(T));
@@ -51,8 +51,8 @@ class LookupTableGradKernel : public framework::OpKernel {
     auto d_output_t = context.Input<Tensor>(framework::GradVarName("Out"));
     auto d_table_t = context.Output<Tensor>(framework::GradVarName("W"));
 
-    size_t N = d_table_t->dims()[0];
-    size_t D = d_table_t->dims()[1];
+    int N = d_table_t->dims()[0];
+    int D = d_table_t->dims()[1];
     auto ids = ids_t->data<int32_t>();
     const T* d_output = d_output_t->data<T>();
     T* d_table = d_table_t->mutable_data<T>(context.GetPlace());
@@ -61,10 +61,10 @@ class LookupTableGradKernel : public framework::OpKernel {
     t.device(context.GetEigenDevice<platform::CPUPlace>()) =
         t.constant(static_cast<T>(0));
 
-    for (size_t i = 0; i < product(ids_t->dims()); ++i) {
+    for (ssize_t i = 0; i < product(ids_t->dims()); ++i) {
       PADDLE_ENFORCE_LT(ids[i], N);
       PADDLE_ENFORCE_GE(ids[i], 0);
-      for (size_t j = 0; j < D; ++j) {
+      for (int j = 0; j < D; ++j) {
         d_table[ids[i] * D + j] += d_output[i * D + j];
       }
     }

paddle/operators/mul_op.cc

@@ -75,8 +75,8 @@ class MulOpGrad : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(y_dims[1] == out_dims[1],
                    "Out@GRAD M X N must equal to Y dims 1, N ");
 
-    x_grad->Resize(x_dims);
-    y_grad->Resize(y_dims);
+    if (x_grad) x_grad->Resize(x_dims);
+    if (y_grad) y_grad->Resize(y_dims);
   }
 };
 

paddle/operators/mul_op.h

@@ -31,13 +31,13 @@ template <typename Place, typename T>
 class MulKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* X = context.Input<Tensor>("X");
-    auto* Y = context.Input<Tensor>("Y");
-    auto* Z = context.Output<Tensor>("Out");
-    Z->mutable_data<T>(context.GetPlace());
+    auto* x = context.Input<Tensor>("X");
+    auto* y = context.Input<Tensor>("Y");
+    auto* z = context.Output<Tensor>("Out");
+    z->mutable_data<T>(context.GetPlace());
     auto* device_context =
         const_cast<platform::DeviceContext*>(context.device_context_);
-    math::matmul<Place, T>(*X, false, *Y, false, 1, Z, 0, device_context);
+    math::matmul<Place, T>(*x, false, *y, false, 1, z, 0, device_context);
   }
 };
 
@@ -45,20 +45,24 @@ template <typename Place, typename T>
 class MulGradKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    auto* X = ctx.Input<Tensor>("X");
-    auto* Y = ctx.Input<Tensor>("Y");
-    auto* dOut = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
 
-    auto* dX = ctx.Output<Tensor>(framework::GradVarName("X"));
-    auto* dY = ctx.Output<Tensor>(framework::GradVarName("Y"));
-    dX->mutable_data<T>(ctx.GetPlace());
-    dY->mutable_data<T>(ctx.GetPlace());
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
     auto* device_context =
         const_cast<platform::DeviceContext*>(ctx.device_context_);
-    // dX = dOut * Y'. dX: M x K, dOut : M x N, Y : K x N
-    math::matmul<Place, T>(*dOut, false, *Y, true, 1, dX, 0, device_context);
-    // dY = X' * dOut. dY: K x N, dOut : M x N, X : M x K
-    math::matmul<Place, T>(*X, true, *dOut, false, 1, dY, 0, device_context);
+    if (dx) {
+      dx->mutable_data<T>(ctx.GetPlace());
+      // dx = dout * y'. dx: M x K, dout : M x N, y : K x N
+      math::matmul<Place, T>(*dout, false, *y, true, 1, dx, 0, device_context);
+    }
+    if (dy) {
+      dy->mutable_data<T>(ctx.GetPlace());
+      // dy = x' * dout. dy K x N, dout : M x N, x : M x K
+      math::matmul<Place, T>(*x, true, *dout, false, 1, dy, 0, device_context);
+    }
   }
 };
 

paddle/operators/rowwise_add_op.cc

@@ -64,8 +64,10 @@ class RowwiseAddGradOp : public framework::OperatorWithKernel {
     auto dims0 = ctx.Input<Tensor>("X")->dims();
     auto dims1 = ctx.Input<Tensor>("b")->dims();
     PADDLE_ENFORCE_EQ(1, dims1.size(), "b dims should be 1")
-    ctx.Output<Tensor>(framework::GradVarName("X"))->Resize(dims0);
-    ctx.Output<Tensor>(framework::GradVarName("b"))->Resize(dims1);
+    auto *dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto *db = ctx.Output<Tensor>(framework::GradVarName("b"));
+    if (dx) dx->Resize(dims0);
+    if (db) db->Resize(dims1);
   }
 };
 

paddle/operators/rowwise_add_op.h

@@ -51,20 +51,24 @@ template <typename Place, typename T>
 class RowwiseAddGradKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* dOut = context.Input<Tensor>(framework::GradVarName("Out"));
-    auto* dX = context.Output<Tensor>(framework::GradVarName("X"));
+    auto* dout = context.Input<Tensor>(framework::GradVarName("Out"));
+    auto* dx = context.Output<Tensor>(framework::GradVarName("X"));
     auto* db = context.Output<Tensor>(framework::GradVarName("b"));
-    dX->mutable_data<T>(context.GetPlace());
-    db->mutable_data<T>(context.GetPlace());
 
-    auto OutGrad = EigenMatrix<T>::From(*dOut);
+    auto out_grad = EigenMatrix<T>::From(*dout);
     auto place = context.GetEigenDevice<Place>();
-    EigenMatrix<T>::From(*dX).device(place) = OutGrad;
+    if (dx) {
+      dx->mutable_data<T>(context.GetPlace());
+      EigenMatrix<T>::From(*dx).device(place) = out_grad;
+    }
 
-    // https://eigen.tuxfamily.org/dox/unsupported/TensorBase_8h_source.html
-    // colwise add
-    Eigen::array<int, 1> dims{{0}}; /* dimension to reduce */
-    EigenVector<T>::Flatten(*db).device(place) = OutGrad.sum(dims);
+    if (db) {
+      db->mutable_data<T>(context.GetPlace());
+      // https://eigen.tuxfamily.org/dox/unsupported/TensorBase_8h_source.html
+      // colwise add
+      Eigen::array<int, 1> dims{{0}}; /* dimension to reduce */
+      EigenVector<T>::Flatten(*db).device(place) = out_grad.sum(dims);
+    }
   }
 };
 }  // namespace operators

python/paddle/trainer/PyDataProvider2.py

@@ -27,6 +27,14 @@ class SequenceType(object):
     SEQUENCE = 1
     SUB_SEQUENCE = 2
 
+    @classmethod
+    def tostring(cls, value):
+        for k in cls.__dict__:
+            if not k.startswith('__'):
+                if getattr(cls, k) == value:
+                    return cls.__name__ + '.' + k
+        return 'INVALID(' + str(value) + ')'
+
 
 # TODO(yuyang18): Add string data type here.
 class DataType(object):
@@ -35,6 +43,14 @@ class DataType(object):
     SparseValue = 2
     Index = 3
 
+    @classmethod
+    def tostring(cls, value):
+        for k in cls.__dict__:
+            if not k.startswith('__'):
+                if getattr(cls, k) == value:
+                    return cls.__name__ + '.' + k
+        return 'INVALID(' + str(value) + ')'
+
 
 class CacheType(object):
     NO_CACHE = 0  # No cache at all
@@ -69,6 +85,26 @@ class InputType(object):
         self.seq_type = seq_type
         self.type = tp
 
+    def __repr__(self):
+        """
+        Return a human readable representation like 'InputType(dim=25921, 
+            seq_type=SequenceType.NO_SEQUENCE, type=DataType.Dense)'
+        """
+        repr_str = type(self).__name__
+        repr_str += '('
+        serialize_func_map = {
+            'dim': repr,
+            'seq_type': SequenceType.tostring,
+            'type': DataType.tostring
+        }
+        for idx, k in enumerate(self.__slots__):
+            if idx != 0:
+                repr_str += ', '
+            repr_str += (
+                k + '=' + serialize_func_map.get(k, repr)(getattr(self, k)))
+        repr_str += ')'
+        return repr_str
+
 
 def dense_slot(dim, seq_type=SequenceType.NO_SEQUENCE):
     """

python/paddle/v2/framework/op.py

@@ -94,9 +94,14 @@ class OpDescCreationMethod(object):
                     new_attr.floats.extend(user_defined_attr)
                 elif attr.type == framework_pb2.STRINGS:
                     new_attr.strings.extend(user_defined_attr)
+                elif attr.type == framework_pb2.INT_PAIRS:
+                    for p in user_defined_attr:
+                        pair = new_attr.pairs.add()
+                        pair.first = p[0]
+                        pair.second = p[1]
                 else:
                     raise NotImplementedError("Not support attribute type " +
-                                              attr.type)
+                                              str(attr.type))
 
         return op_desc
 

python/paddle/v2/framework/tests/gradient_checker.py

@@ -286,6 +286,9 @@ class GradientChecker(unittest.TestCase):
         for no_grad in no_grad_set:
             if no_grad not in in_names:
                 raise ValueError("no_grad should be in in_names")
+            if no_grad in inputs_to_check:
+                raise ValueError("no_grad should not be in inputs_to_check")
+
         backward_op = core.Operator.backward(forward_op, no_grad_set)
 
         places = [core.CPUPlace()]
@@ -301,7 +304,6 @@ class GradientChecker(unittest.TestCase):
 
         check_names = [grad_var_name(name) for name in inputs_to_check]
         for place in places:
-            # get analytical gradients according to different device
             analytic_grads = self.__get_gradient(forward_op, backward_op,
                                                  input_vars, check_names, place)
             self.__assert_is_close(numeric_grads, analytic_grads, check_names,

python/paddle/v2/framework/tests/test_fc_op.py

@@ -2,6 +2,7 @@ import unittest
 import numpy as np
 from gradient_checker import GradientChecker, create_op
 from op_test_util import OpTestMeta
+from paddle.v2.framework.op import Operator
 
 
 class TestFCOp(unittest.TestCase):
@@ -18,12 +19,23 @@ class TestFCOp(unittest.TestCase):
         mul_out = np.dot(self.inputs["X"], self.inputs["W"])
         add_out = np.add(mul_out, self.inputs["b"])
         sigmoid_out = 1 / (1 + np.exp(-add_out))
-        self.outputs = {"mul_out": add_out, "Y": sigmoid_out}
+        self.outputs = {
+            "mul_out": mul_out,
+            "add_out": add_out,
+            "Out": sigmoid_out
+        }
 
 
 class TestFCGradOp(GradientChecker):
     def test_normal(self):
-        print "nothing"
+        self.inputs = {
+            "X": np.random.random((4, 4)).astype("float32"),
+            "W": np.random.random((4, 4)).astype("float32"),
+            "b": np.random.random(4).astype("float32")
+        }
+        op = Operator(
+            "fc", X="X", W="W", b="b", Out="Out", activation="sigmoid")
+        #self.check_grad(op, self.inputs, ["X", "W", "b"], "Out")
 
 
 if __name__ == '__main__':

python/paddle/v2/framework/tests/test_mul_op.py

@@ -16,16 +16,37 @@ class TestMulOp(unittest.TestCase):
         self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
 
 
-class MulGradOpTest(GradientChecker):
-    def test_mul(self):
-        op = create_op("mul")
-        inputs = {
+class TestMulGradOp(GradientChecker):
+    def setUp(self):
+        self.op = create_op("mul")
+        self.inputs = {
             'X': np.random.random((32, 84)).astype("float32"),
             'Y': np.random.random((84, 100)).astype("float32")
         }
+
+    def test_cpu_gpu_compare(self):
+        self.compare_grad(self.op, self.inputs)
+
+    def test_normal(self):
         # mul op will enlarge the relative error
         self.check_grad(
-            op, inputs, set(["X", "Y"]), "Out", max_relative_error=0.5)
+            self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.5)
+
+    def test_ignore_x(self):
+        self.check_grad(
+            self.op,
+            self.inputs, ["Y"],
+            "Out",
+            max_relative_error=0.5,
+            no_grad_set={"X"})
+
+    def test_ignore_y(self):
+        self.check_grad(
+            self.op,
+            self.inputs, ["X"],
+            "Out",
+            max_relative_error=0.5,
+            no_grad_set={"Y"})
 
 
 # TODO(dzh,qijun) : mulgrad test case need transpose feature of blas library

python/paddle/v2/framework/tests/test_rowwise_add_op.py

@@ -16,14 +16,22 @@ class TestRowwiseAddOp(unittest.TestCase):
         self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['b'])}
 
 
-class RowwiseAddGradOpTest(GradientChecker):
-    def test_rowwise_add(self):
-        op = create_op("rowwise_add")
-        inputs = {
+class TestRowwiseAddGradOp(GradientChecker):
+    def setUp(self):
+        self.op = create_op("rowwise_add")
+        self.inputs = {
             "X": np.random.uniform(0.1, 1, [5, 10]).astype("float32"),
             "b": np.random.uniform(0.1, 1, [10]).astype("float32")
         }
-        self.check_grad(op, inputs, set(["X", "b"]), "Out")
+
+    def test_normal(self):
+        self.check_grad(self.op, self.inputs, ["X", "b"], "Out")
+
+    def test_ignore_b(self):
+        self.check_grad(self.op, self.inputs, ["X"], "Out", no_grad_set={"b"})
+
+    def test_ignore_x(self):
+        self.check_grad(self.op, self.inputs, ["b"], "Out", no_grad_set={"X"})
 
 
 if __name__ == '__main__':